AI300
  1. Noam Brown / OpenAI Research Scientist
    1801.
    @offchan420 Yes but the point is specifying the reward in a way that leads pure self play to optimal real-world behavior is virtually impossible.
    LLM score 80 · 7 months ago
  2. Naman Jain / Researcher at Cursor
    1802.
    We added LLM judge based hack detector to our code optimization evals and found models perform non-idiomatic code changes in upto 30% of the problems 🤯 https://t.co/CPwEQvRQJC
    LLM score 85 · 7 months ago
  3. Andrej Karpathy / AI researcher
    1803.
    @GoodFaithOnly @binarybits When I say AGI I mean a "feature complete" build of a remote worker with a lot of 9s (think: a bit like current Autopilot FSD, maybe plus a few more iterations).›
    LLM score 70 · 7 months ago
  4. Emmanuel Ameisen / Anthropic Interpretability Researcher
    1804.
    Neat use of attribution graphs to identify features in models which are responsible for jailbreak resistance and refusals.›
    LLM score 85 · 7 months ago
  5. Hang Gao / ex MTS at xAI
    1805.
    The value of eval lies in a distributional look at the model performance.›
    LLM score 92 · 7 months ago
  6. Victoria Lin / Thinking Machines Researcher
    1806.
    Very interesting read ☕ When poking different frontier models (e.g., GPT-5 vs Gemini), I’ve often noticed surprising similarity on non-STEM questions.›
    LLM score 85 · 7 months ago
  7. Xiang Fu / Researcher at Periodic Labs
    1807.
    Second @simonbatzner. FF architecture was no longer a bottleneck for AI materials discovery since nequip.›
    LLM score 80 · 7 months ago
  8. Noam Brown / OpenAI Research Scientist
    1808.
    @lineardiff @ericzelikman I don’t view that as incompatible.›
    LLM score 20 · 7 months ago
  9. Emmanuel Ameisen / Anthropic Interpretability Researcher
    1809.
    Striking result, which changed how I think about LLMs:›
    LLM score 85 · 7 months ago
  10. Niklas Muennighoff / AI Researcher at Stanford
    1810.
    new paper "ATLAS" – for details: https://t.co/MJO9mvlS90; it also covers a new scaling law, pretrain vs finetune for multilingual & more :)›
    LLM score 70 · 7 months ago
  11. Niklas Muennighoff / AI Researcher at Stanford
    1811.
    To scale data-constrained LLMs, repeating & denoising objectives can help.›
    LLM score 85 · 7 months ago
  12. Tri Dao / Chief Scientist at Together
    1812.
    State space architecture for state of the art voice model! https://t.co/QjdOdrIFSM
    LLM score 80 · 7 months ago
  13. Emmanuel Ameisen / Anthropic Interpretability Researcher
    1813.
    Cool paper using attribution graphs to automatically detect which reasoning steps contain mistakes!›
    LLM score 85 · 7 months ago
  14. Lilian Weng / Thinking Machines Cofounder
    1814.
    On-policy distillation provides an elegant way to use the teacher model as a process reward model to provide dense reward while preventing SFT style "OOD shock" during rollout.›
    LLM score 90 · 7 months ago
  15. Mira Murati / Thinking Machines CEO
    1815.
    Combining the benefits of RL and SFT with on-policy distillation, a promising approach for training small models for domain performance and continual learning.›
    LLM score 85 · 7 months ago
  16. Xiang Fu / Researcher at Periodic Labs
    1816.
    To adapt to today's world, crank down discount factor and crank up learning rate
    LLM score 75 · 7 months ago
  17. John Schulman
    1817.
    Happy to share a new paper! Designing model behavior is hard -- desirable values often pull in opposite directions.›
    LLM score 80 · 7 months ago
  18. Xiang Fu / Researcher at Periodic Labs
    1818.
    Every AI researcher should read two essays.›
    LLM score 80 · 7 months ago
  19. Sally Zhu / Researcher at Flapping Airplanes
    1819.
    🔎Did someone steal your language model? We can tell you, as long as you shuffled your training data🔀.›
    LLM score 95 · 7 months ago
  20. Aleksander Madry / OpenAI Researcher
    1820.
    Does removing a data source improve model capabilities? For instance, does adding another language data source increase model toxicity?›
    LLM score 75 · 7 months ago
  21. Noam Brown / OpenAI Research Scientist
    1821.
    Below is a deep dive into why self play works for two-player zero-sum (2p0s) games like Go/Poker/Starcraft but is so much harder to use in "real world" domains.›
    LLM score 80 · 8 months ago
  22. Noam Brown / OpenAI Research Scientist
    1822.
    @pli_cachete @karpathy Because it means self play converges to a minimax equilibrium which is essentially an unbeatable strategy.›
    LLM score 80 · 8 months ago
  23. Noam Brown / OpenAI Research Scientist
    1823.
    Self play works so well in chess, go, and poker because those games are two-player zero-sum.›
    LLM score 75 · 8 months ago
  24. Noam Brown / OpenAI Research Scientist
    1824.
    .@Stanford courses are high-quality but the policies are definitely outdated.›
    LLM score 20 · 8 months ago
  25. Demis Hassabis / CEO of DeepMind
    1825.
    Awesome to see Veo 3.1 top the LMArena video leaderboards by a large distance with big improvements over Veo 3.0 for text-to-video (+30) and image-to-video (+70)! 🔥Huge congrats to the team! Try it for yourself in https://t.co/QgTpxTL8DQ and the @GeminiApp https://t.co/nQ14ZnkKnr
    LLM score 20 · 8 months ago
  26. Noam Brown / OpenAI Research Scientist
    1826.
    @fchollet Didn't you say just two months ago that you think AGI is about 5 years away?›
    LLM score 20 · 8 months ago
  27. Richard Song / DeepMind Research Scientist
    1827.
    Add-on: from our experience with Regression Language Models which absorb evaluations via weight updates - it takes at least *thousands* of (x,y) pairs to learn any meaningful world model of the objective.›
    LLM score 92 · 8 months ago
  28. Demis Hassabis / CEO of DeepMind
    1828.
    Super excited to be collaborating with Commonwealth Fusion Systems @CFS_energy to use AI to accelerate fusion development - and move closer to a sustainable future with limitless clean energy https://t.co/kCS1UKEtjP
    LLM score 70 · 8 months ago
  29. Demis Hassabis / CEO of DeepMind
    1829.
    Veo 3 is the state-of-the-art in video models.›
    LLM score 30 · 8 months ago
  30. Demis Hassabis / CEO of DeepMind
    1830.
    We processed over 1.3 Quadrillion tokens last month - that's 1,300,000,000,000,000 tokens! or to put it another way that's 500M tokens a second or 1.8 Trillion tokens an hour...›
    LLM score 75 · 8 months ago
More